Diogo – LLM, RAG, LangChain

A legal tech platform that predicts how a court case is likely to unfold. General-purpose LLMs couldn’t reliably forecast this, and they performed no better than an informed guess. The platform models each case’s event history together with structured context such as jurisdiction and case subject. It uses this to forecast next events and likely outcomes for legal professionals. Its models are trained on a historical dataset of 30M+ case records.

• Leads end-to-end development of the platform’s ML system. Builds custom transformer models that outperform frontier LLMs by over 3x on key legal prediction tasks, results that helped the company secure multi-million investment.
• Designed the modeling approach through measured iterations: from prompting frontier LLMs, to fine-tuning open models with QLoRA, to a BERT-based classifier that embeds jurisdiction and case subject as categorical features alongside case event history.
• Set up entropy-based baselines and leakage controls (case-wise data splits, temporal masking of future events) so reported gains reflect real predictive power.
• Builds reproducible training and evaluation pipelines with PyTorch, Hugging Face, MLflow and Docker, and hands off production-ready models for deployment.
• Runs EDA and ETL over 30M+ rows using SQL and Ray to prepare training data.
• Guides supporting contributors and reports progress to leadership through weekly executive summaries.

A job market intelligence platform that turns large volumes of job postings data into insights for analysts and business stakeholders. It replaces manual data labeling and ad hoc analysis. New job market records are ingested and classified every month and stored in a data warehouse. Stakeholders explore the data through interactive dashboards and a conversational AI assistant that answers questions in plain language. The database holds 100K+ records and grows by 1,500+ records monthly.

• Built and deployed a chat-based AI agent with LangChain, RAG, Milvus and PostgreSQL that lets stakeholders query a 100K+ row job market database in plain language, cutting time to insight by over 30%.
• Built a generative AI data labeling tool on LLM APIs (OpenAI) and Haystack that classifies records with 86% accuracy, saving 300+ working hours per month.
• Designed and deployed an ETL pipeline with Python, Airflow and AWS (EC2, S3) that automates ingestion into the data warehouse, adding 1,500+ new records monthly.
• Traced and evaluated LLM pipelines with LangSmith.
• Developed a predictive modeling app and Metabase dashboards with Plotly visualizations for stakeholder reporting.

A SaaS platform that generates analytical performance reports for content creators. Before it, a report took around two months to prepare manually. The platform collects creators’ data into a managed database and runs machine learning and generative AI analysis over it. It then delivers a ready-to-read report with visualizations.

• Built the SaaS end to end, automating analytical report generation and cutting delivery time from 60 days to 5 minutes.
• Designed the database and backend with Supabase and FastAPI.
• Integrated LLM APIs and machine learning models to generate the report’s analysis and insights.
• Built the visualization layer and deployed the application to the cloud.

Psychometric assessments developed in a university research lab to measure different dimensions of human behavior. The lab designs new tests, validates them statistically on participant data, and publishes the findings in peer-reviewed journals.

• Co-developed and validated 3 original psychological tests using advanced statistical modeling and machine learning in R.
• Contributed to 23 peer-reviewed scientific publications.
• Processed and visualized research data to support test validation and publication.